如何将Keras模型权重转换为一维数组？

问题背景 #

基于Keras构建的InceptionV3模型处理MNIST数据集的代码：

x_train_mnist = np.expand_dims(x_train_mnist, axis=-1)     
x_train_mnist = np.repeat(x_train_mnist, 3, axis=-1)        

x_train_mnist = x_train_mnist.astype('float32') / 255

x_train_mnist = tf.image.resize(x_train_mnist, [75,75]) 

x_test_mnist = np.expand_dims(x_test_mnist, axis=-1)

x_test_mnist = np.repeat(x_test_mnist, 3, axis=-1)         

x_test_mnist = x_test_mnist.astype('float32') / 255

x_test_mnist = tf.image.resize(x_test_mnist, [75,75])

model=tf.keras.applications.InceptionV3(
 include_top=True,
    pooling=None,
    classes=10,
    weights=None,
    input_shape=(75,75,3)
) 

尝试转换权重为一维数组时执行以下代码出错：

weights=model.weights
print(type(weights))
weights_np=np.array(weights)
print(type(weights_np))
weights_npfloat_32=np.float32(weights_np)
print(weights_np.size) 

错误输出：

<class 'list'>
<class 'numpy.ndarray'>
/tmp/ipykernel_3796671/1208296518.py:3: VisibleDeprecationWarning: Creating an ndarray from ragged nested sequences (which is a list-or-tuple of lists-or-tuples-or ndarrays with different lengths or shapes) is deprecated. If you meant to do this, you must specify 'dtype=object' when creating the ndarray.
  weights_np=np.array(weights)
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
File ~/miniconda3/lib/python3.9/site-packages/tensorflow/python/ops/resource_variable_ops.py:1456, in BaseResourceVariable.__float__(self)
   1455 def __float__(self):
-> 1456   return float(self.value().numpy())

TypeError: only size-1 arrays can be converted to Python scalars

The above exception was the direct cause of the following exception:

ValueError                                Traceback (most recent call last)
Cell In [42], line 5
      3 weights_np=np.array(weights)
      4 print(type(weights_np))
----> 5 weights_npf=np.float32(weights_np)
      6 print(weights_np.size)

ValueError: setting an array element with a sequence.

错误原因 #

• model.weights返回的是TensorFlow ResourceVariable对象的列表，并非直接的numpy数组，直接转换会生成存储对象的numpy数组，无法直接转为float32类型
• 每个权重变量的形状各不相同，直接用np.array(weights)会生成不规则数组（ragged array），触发警告且后续类型转换失败

解决方案 #

需要先将每个权重变量转为numpy数组并展平，再合并为一个一维数组：

import numpy as np

# 遍历所有权重，逐个转换为numpy数组并展平
flattened_weights = []
for weight_var in model.weights:
    # 将TensorFlow变量转为numpy数组，再展平为一维
    weight_np = weight_var.numpy().flatten()
    flattened_weights.append(weight_np)

# 合并所有一维数组为一个大的行向量
weights_1d = np.concatenate(flattened_weights)

# 验证结果
print(f"类型：{type(weights_1d)}")
print(f"形状：{weights_1d.shape}")  # 输出形如(总权重数量,)的一维数组
print(f"数据类型：{weights_1d.dtype}")  # 应为float32

代码说明 #

1. 遍历model.weights中的每个变量，用weight_var.numpy()将TensorFlow变量转为numpy数组，再通过flatten()将其转为一维数组
2. 使用np.concatenate将所有一维数组合并成一个连续的行向量，最终得到包含所有模型权重的一维数组

内容的提问来源于stack exchange，提问作者CA Khan